Facilitating high-dimensional transparent classification via empirical Bayes variable selection

摘要

We present a two-step approach to classification problems in the large P, small N setting, where the number of predictors may be larger than the sample size. We assume that the association between the predictors and the class variable has an approximate linear-logistic form, but we allow the class boundaries to be nonlinear. We further assume that the number of true predictors is relatively small. In the first step, we use a binomial generalized linear model to identify which predictors are associated with each class and then restrict the data set to these predictors and run a nonlinear classifier, such as a random forest or a support vector machine. We show that, without the variable screening step, the classification performance of both the random forest and support vector machine is degraded when many among the P predictors are not related to the class.